LAUSR

Increasing availability of medico-administrative databases has prompted the development of automated pharmacovigilance signal detection methodologies. Self-controlled approaches have recently been proposed. They account for time-independent confounding factors that may not be recorded. So far, large numbers of drugs have been screened either univariately or with LASSO penalized regressions. We propose and assess a new method that combines the case-crossover self-controlled design with propensity scores (propensity score-adjusted case-crossover) built from high-dimensional data-driven variable selection, to account for co-medications or possibly other measured confounders. Comparison with the univariate and LASSO case-crossover was performed from simulations and a real-data study. Multiple regressions (LASSO, propensity score-adjusted case-crossover) accounted for co-medications and no other covariates. For the univariate and propensity score-adjusted case-crossover methods, the detection threshold was based on a false discovery rate procedure, while for LASSO, it relied on the Akaike Information Criterion. For the real-data study, two drug safety experts evaluated the signals generated from the analysis of 4099 patients with acute myocardial infarction from the French national health database. On simulations, our approach ranked the signals similarly to the LASSO and better than the univariate method while controlling the false discovery rate at the prespecified level, contrary to the univariate method. The LASSO provided the best sensitivity at the cost of larger false discovery rate estimates. On the application, our approach showed similar performances to the LASSO and better performances than the univariate method. It highlighted 43 signals out of 609 drug candidates: 22 (51%) were considered as potentially pharmacologically relevant, including seven (16%) regarded as highly relevant. Our findings show the interest of a propensity score combined with a case-crossover for pharmacovigilance. They also confirm that indication bias remains a challenge when mining medico-administrative databases.